Conventional configuration techniques for three-dimensional (3D) workloads are incapable of executing optimal configuration of 3D workloads for any number of reasons, such as the entire previous and next time windows are not known for executing the same type of workload. For example, conventional configuration techniques are not capable of knowing optimal machine configurations until after workloads have been executed and such conventional techniques average measurements over a time window that includes different types of work, resulting in a machine configuration that may be acceptable for average characteristics of that work, but not optimal for any specific subslice of it and further, such a configuration is incapable of being applied to previous chunks of work.